Paths Toward AGI and Post-Transformer Intelligence

The preceding chapters established the architectural foundations of DOGMA: its six-stage pipeline (Chapter 7), the Tera regime dynamics (Chapter 9), the tetrahedral data substrate (Chapter 10), and the scaling arguments that position it beyond the transformer paradigm (Chapter 15). Each contribution addressed a specific engineering challenge. This chapter asks the larger question: what does it mean to build artificial general intelligence, and what architectural paths might lead there?

We proceed carefully. AGI is a term laden with hype, ambiguity, and premature claims. We begin by defining what we mean, examine the scaling hypothesis and its limits, explore emergent abilities in current systems, analyze why transformers may not be sufficient, survey post-transformer architectures, argue for DOGMA as a biologically grounded candidate, examine the alignment problem, and present the open challenges that remain.

16.1 What Is Artificial General Intelligence?

16.1.1 Definitions and the Capability Spectrum

No universally accepted definition of AGI exists. The term was popularized by Goertzel and Pennachin [2014] to distinguish systems with broad, flexible intelligence from the narrow AI systems that dominate commercial applications. We adopt a working definition that emphasizes generality over performance:

This definition positions AGI on a capability spectrum rather than as a binary threshold. To make this spectrum concrete, consider the following levels of AI capability, loosely inspired by frameworks from Morris et al. [2023]:

Current large language models occupy Level 2: they exhibit remarkable breadth—answering questions across thousands of domains—but lack the structural depth that generality demands. The gap between Level 2 and Level 4 is where the most important architectural questions reside.

16.1.2 Narrow AI vs. General AI: A Comparison

The distinction between narrow and general AI is not merely quantitative (more tasks) but qualitative (different kind of competence):

16.1.3 Why Current LLMs Are Broad but Not General

Despite their fluency, LLMs fail several criteria in Definition 16.1.1:

1.

No persistent state. An LLM’s context window resets between sessions. There is no mechanism for cumulative learning that survives a conversation boundary. Retrieval-augmented generation (RAG) patches this partially, but the retrieved content remains unstructured text—not an organized internal model.

2.

Shallow reasoning. Chain-of-thought prompting [Wei et al., 2022b] elicits reasoning traces, but these are generated autoregressively without the ability to backtrack, verify intermediate steps, or maintain a structured proof state. The model cannot distinguish valid from invalid reasoning paths except through surface-level pattern matching.

3.

No world model. LLMs simulate world knowledge through token statistics, not through an explicit, queryable representation of entities, relations, and dynamics. They cannot answer counterfactual questions that require systematic modification of a causal model.

4.

No autonomy. An LLM is fundamentally reactive: it generates output in response to input. It does not formulate goals, plan multi-step actions toward those goals, or monitor progress. Agentic scaffolding [Yao et al., 2023] provides some autonomy, but the scaffolding is external, not intrinsic to the architecture.

16.2 The Scaling Hypothesis

16.2.1 What the Scaling Hypothesis Claims

The scaling hypothesis holds that increasing model size, dataset size, and compute is sufficient to produce increasingly capable—and eventually general—intelligence. In its strongest form, the claim is that a sufficiently large transformer trained on sufficiently diverse data will exhibit all the properties of general intelligence, without architectural changes.

This hypothesis draws support from the remarkably consistent scaling laws observed in large language models. Kaplan et al. [2020b] demonstrated that the cross-entropy loss \(L\) of a language model follows power-law relationships with the number of parameters \(N\), the dataset size \(D\), and the compute budget \(C\):

\begin {equation} L(N) \approx \left (\frac {N_c}{N}\right )^{\alpha _N}, \quad L(D) \approx \left (\frac {D_c}{D}\right )^{\alpha _D}, \quad L(C) \approx \left (\frac {C_c}{C}\right )^{\alpha _C} \label {eq:ch16-scaling-laws} \end {equation}

where \(\alpha _N \approx 0.076\), \(\alpha _D \approx 0.095\), and \(\alpha _C \approx 0.050\) are empirically measured exponents, and \(N_c, D_c, C_c\) are constants.

16.2.2 Chinchilla and Compute-Optimal Training

An important refinement came from Hoffmann et al. [2022b], who demonstrated that most large language models are significantly undertrained. The Chinchilla scaling laws showed that for a fixed compute budget \(C\), the optimal allocation between parameters \(N\) and training tokens \(D\) is approximately:

\begin {equation} N_{\text {opt}} \propto C^{0.5}, \quad D_{\text {opt}} \propto C^{0.5} \label {eq:ch16-chinchilla} \end {equation}

This means parameters and data should scale equally with compute. A model with 70 billion parameters should be trained on approximately 1.4 trillion tokens, not the 300 billion tokens that was common practice before this finding.

16.2.3 Arguments For the Scaling Hypothesis

Proponents of the scaling hypothesis point to several compelling observations:

1.

Consistent improvement. Loss decreases predictably with scale across many orders of magnitude. No “wall” has been observed where scaling stops helping.

2.

Emergent capabilities. Larger models exhibit qualitatively new behaviors (Section 16.3) that smaller models cannot perform at all.

3.

Cross-domain generalization. Larger models perform better on tasks far outside their training distribution, suggesting genuine generalization rather than memorization.

4.

Universality. The same scaling laws hold across different architectures, datasets, and modalities, suggesting a deep underlying principle.

16.2.4 Arguments Against the Scaling Hypothesis

Critics raise equally compelling objections:

1.

Loss is not capability. Scaling laws describe loss (perplexity on held-out text), not task-level capability. A model can have low perplexity but still fail at basic reasoning tasks.

2.

Data wall. High-quality text data is finite. Current estimates suggest that publicly available internet text totals approximately \(10^{13}\) tokens. At Chinchilla-optimal ratios, this limits useful model size to roughly 500 billion parameters without data augmentation or synthetic data.

3.

Diminishing returns. While loss continues to decrease, the rate of improvement slows. The exponents in Equation 16.1 are small (\(< 0.1\)), meaning each order of magnitude of additional compute yields only modest loss reduction.

4.

Fundamental capability gaps. Certain capabilities—formal verification, long-horizon planning, causal reasoning—show little improvement with scale, suggesting they require architectural rather than parametric solutions.

5.

Economic infeasibility. Training a model 1000\(\times \) larger than current frontier models would cost hundreds of billions of dollars in compute alone, not counting energy, cooling, and infrastructure.

16.3 Emergent Abilities in Large Models

16.3.1 What Is Emergence?

An ability is called emergent if it is absent in smaller models and appears—seemingly abruptly—as the model scales beyond a critical threshold. The term, borrowed from complexity science, captures the idea that quantitative changes (more parameters) can produce qualitative changes (new capabilities).

16.3.2 Step-by-Step Examples of Emergence

We examine three prominent emergent abilities in detail.

Example 1: In-Context Learning. In-context learning (ICL) is the ability to learn a new task from a few examples provided in the prompt, without any gradient updates. Consider a sentiment classification task:

Small models (\(< 1\)B parameters) treat these examples as unrelated text and generate random continuations. Models above approximately 6B parameters begin to extract the pattern and correctly output “Negative.” Models above 100B parameters can perform ICL on much more complex tasks, including arithmetic, translation between rare languages, and logical deduction.

The mechanism behind ICL remains debated. Olsson et al. [2022] identified “induction heads”—attention head circuits that implement a copy-and-complete pattern—as a key mechanism. But why these circuits emerge only at sufficient scale is not fully understood.

Example 2: Chain-of-Thought Reasoning. Chain-of-thought (CoT) prompting [Wei et al., 2022b] instructs the model to show its reasoning step by step before producing a final answer:

CoT is emergent: models below approximately 60B parameters show no benefit from the “Let’s think step by step” prompt, while larger models show dramatic improvement on arithmetic, symbolic reasoning, and commonsense reasoning tasks.

Example 3: Tool Use. The ability to recognize when an external tool (calculator, search engine, code interpreter) would help and generate the correct API call to invoke it is another emergent capability. A model might generate:

This requires the model to (1) recognize the limits of its own knowledge, (2) identify the appropriate tool, and (3) formulate a well-structured query—a combination that only emerges at large scale.

16.3.3 Are Emergent Abilities Real?

Schaeffer et al. [2023b] argued that emergent abilities may be a measurement artifact: they appear abrupt only because researchers use discontinuous metrics (exact-match accuracy) rather than continuous ones (log-likelihood). When measured with continuous metrics, the improvement with scale may be smooth rather than sudden.

This debate highlights an important distinction: emergent performance (abrupt improvement on a benchmark) versus emergent capability (a qualitatively new kind of behavior). Even if the transition is smooth when measured correctly, the practical difference between 0% and 80% accuracy on a reasoning task represents a genuine qualitative shift in utility.

16.4 Why Transformers May Not Be Enough

The transformer architecture [Vaswani et al., 2017b] has dominated AI since 2017. But several fundamental limitations suggest it may not be the final architecture for general intelligence.

16.4.1 The Quadratic Attention Bottleneck

Self-attention computes all-pairs interactions in \(O(T^2)\) time and memory:

\begin {equation} \text {Attention}(\mathbf {Q}, \mathbf {K}, \mathbf {V}) = \text {softmax}\!\left (\frac {\mathbf {Q}\mathbf {K}^\top }{\sqrt {d_k}}\right )\mathbf {V} \label {eq:attention-reminder} \end {equation}

For a sequence of length \(T\) with model dimension \(d\), this requires \(O(T^2 \cdot d)\) FLOPs per layer. The practical consequence is severe:

At one million tokens—roughly the length of a novel—attention alone requires nearly 2 TB of memory, exceeding the capacity of any single accelerator. Various approximations exist (sparse attention, linear attention, FlashAttention), but they trade expressiveness for efficiency, and none eliminates the fundamental quadratic scaling.

16.4.2 Reasoning Limitations

Transformers process information in a fixed number of forward passes (one per layer). This means the computational depth available for reasoning is determined at design time, not at inference time. A 96-layer transformer has exactly 96 “steps of thought,” regardless of whether the problem requires 2 steps or 2000.

Furthermore, Merrill and Sabharwal [2023] showed that fixed-precision transformers with \(O(1)\) layers can recognize only languages in the complexity class \(\mathsf {TC^0}\) (constant-depth threshold circuits). Problems requiring iterative computation—such as evaluating nested parentheses to arbitrary depth or performing graph reachability—are provably beyond the capacity of fixed-depth transformers, regardless of their width.

16.4.3 The World Model Problem

A world model is an internal representation that captures the structure of the environment and supports counterfactual reasoning (“what would happen if…”). Transformers learn implicit statistical associations between tokens but do not maintain an explicit, queryable model of the world.

Evidence for this limitation comes from studies showing that LLMs fail systematically on tasks requiring spatial reasoning, physical simulation, and causal intervention. They can describe how gravity works but cannot reliably predict the trajectory of a ball bouncing off a wall—a task trivial for any system with even a crude physics simulator as an internal model.

16.5 Post-Transformer Architectures

Several architectural families have emerged as candidates to address the transformer’s limitations. We survey the most promising.

16.5.1 State Space Models: Mamba and Beyond

State space models (SSMs) replace attention with a linear recurrence derived from continuous-time dynamical systems. The Mamba architecture [Gu and Dao, 2023] introduced selective state spaces, where the state transition matrices are input-dependent:

\begin {equation} h_t = \bar {A}_t \, h_{t-1} + \bar {B}_t \, x_t, \quad y_t = C_t \, h_t \label {eq:ssm} \end {equation}

where \(h_t \in \mathbb {R}^n\) is the hidden state, \(x_t\) is the input, and \(\bar {A}_t, \bar {B}_t, C_t\) are discretized, input-dependent matrices. The key advantages are:

• Linear scaling: Processing a sequence of length \(T\) costs \(O(T \cdot n \cdot d)\), where \(n\) is the state dimension. There is no quadratic term.
• Efficient training: The recurrence can be computed via a parallel scan algorithm, enabling training throughput comparable to transformers.
• Infinite context: The fixed-size state \(h_t\) compresses the entire history, enabling theoretically infinite context without growing memory.

However, SSMs also have limitations. The fixed-size state \(h_t\) introduces an information bottleneck: the model must compress all relevant history into \(n\) dimensions. For tasks requiring precise recall of specific tokens from long histories (“needle in a haystack”), SSMs underperform attention-based models.

16.5.2 Mixture of Experts

Mixture-of-Experts (MoE) architectures activate only a subset of the model’s parameters for each input, achieving large effective model capacity with tractable compute costs. In a typical MoE layer:

\begin {equation} y = \sum _{i=1}^{E} g_i(x) \cdot \text {Expert}_i(x), \quad g(x) = \text {TopK}\bigl (\text {softmax}(W_g \, x)\bigr ) \label {eq:ch16-moe} \end {equation}

where \(E\) is the number of experts (typically 8–64), \(g_i(x)\) is a gating function that selects the top-\(K\) experts (typically \(K = 2\)), and each expert is an independent feed-forward network.

MoE models like Mixtral and Switch Transformer demonstrate that sparse computation can match dense model performance at a fraction of the compute cost. However, MoE addresses only the compute wall; it does not fundamentally change the attention mechanism, reasoning depth, or world modeling capacity.

16.5.3 Liquid Neural Networks

Liquid neural networks, inspired by the nervous systems of small organisms like C. elegans, use continuous-time differential equations with time-varying parameters:

\begin {equation} \frac {dh}{dt} = -\frac {h}{\tau (x)} + f(h, x, t; \theta ) \label {eq:liquid} \end {equation}

where \(\tau (x)\) is an input-dependent time constant that controls how quickly the network responds to new information. The key property is adaptive computation: the network automatically allocates more processing time to complex inputs and less to simple ones.

Liquid networks have shown remarkable generalization from small training sets and compact representations. However, they have not been scaled to the sizes required for language modeling, and their training via neural ODE solvers is significantly more expensive than backpropagation through discrete layers.

16.5.4 Neuromorphic Computing

Neuromorphic architectures, such as Intel’s Loihi and IBM’s TrueNorth chips, implement spiking neural networks directly in hardware. Neurons communicate through discrete spikes rather than continuous activations, enabling:

• Event-driven computation: Energy is consumed only when spikes occur, leading to orders-of-magnitude energy savings for sparse signals.
• Temporal coding: Information is encoded in spike timing, not just spike rates, providing a richer representational substrate.
• Local learning rules: Spike-timing-dependent plasticity (STDP) enables learning without global backpropagation.

Neuromorphic systems excel at temporal pattern recognition and sensorimotor tasks but currently lack the representational capacity and training methodology to compete with transformers on language and reasoning tasks. Their primary contribution to the AGI path may be as co-processors for specific capabilities (real-time perception, energy-efficient inference) rather than as complete architectures.

16.6 DOGMA as a Post-Transformer Candidate

Given the AGI criteria in Definition 16.1.1 and the limitations of existing architectures, we now examine how DOGMA’s biologically inspired design addresses each of the transformer’s three walls.

16.6.1 Addressing the Compute Wall: Regulated Propagation

DOGMA replaces global attention with regulated propagation: information flows through the tetrahedral mesh (Chapter 10) along paths determined by the regulatory network. The computational cost scales as:

\begin {equation} C_{\text {DOGMA}}(T) = O(T \cdot k \cdot d) \label {eq:dogma-cost} \end {equation}

where \(k\) is the average number of active neighbors in the tetrahedral mesh and \(d\) is the propagation depth needed for global information flow. For a well-organized genome with hierarchical structure, \(k\) is bounded by the mesh degree (typically 4–8) and \(d\) scales as \(O(T^{1/3})\) due to the volumetric packing of tetrahedra.

\begin {equation} C_{\text {DOGMA}}(T) = O(T^{4/3} \cdot k) \quad \text {vs.} \quad C_{\text {Transformer}}(T) = O(T^2 \cdot d_{\text {model}}) \label {eq:cost-comparison} \end {equation}

16.6.2 Addressing the Reasoning Wall: Dynamic Regulation Depth

Unlike transformers, which process every input through the same fixed number of layers, DOGMA’s regulatory network can activate different computational pathways for different inputs. Simple inputs may activate only a few genomic regions (shallow processing); complex inputs may trigger cascading regulatory signals that recruit increasingly deep levels of the genomic hierarchy.

Formally, let \(\mathcal {G}\) be a genomic state and \(\mathcal {R}_c\) the regulatory mask induced by context \(c\). The number of active computation stages is determined dynamically:

\begin {equation} \text {Depth}(c) = \max \bigl \{ i : \|\mathcal {R}_c^{(i)}\|_0 > 0 \bigr \} \label {eq:dynamic-depth} \end {equation}

This provides the computational analogue of what cognitive science calls adaptive resource allocation: spending more cognitive effort on harder problems.

16.6.3 Addressing the Grounding Wall: Structured State as World Model

DOGMA’s genomic structure provides typed, organized internal state (Chapter 7). Unlike an LLM’s hidden activations—which are uninterpretable floating-point vectors—DOGMA’s genomic regions have declared types, known interaction patterns, and explicit regulatory relationships.

This structure can serve as a world model: entities are represented as genomic regions, relationships as regulatory connections, and dynamics as regulatory cascades. Counterfactual reasoning becomes structural modification: “what would happen if…” is implemented by modifying a regulatory signal and observing how the cascade propagates through the genome.

\begin {equation} \text {Verify}(\mathcal {G}, c, y) = \bigwedge _{i=1}^{n} \text {TypeCheck}\bigl (\mathcal {R}_c^{(i)}, \mathcal {G}^{(i)}, \mathcal {T}^{(i)}\bigr ) \label {eq:verify} \end {equation}

where \(\mathcal {T}^{(i)}\) is the expected type signature at stage \(i\) and the conjunction ensures that every active region satisfies its type constraints.

16.6.4 Compositional Verification Through Tetrahedral Structure

TetraData (Chapter 10) provides a volumetric data structure that enables compositional verification. Each tetrahedron encodes four related quantities with six pairwise relationships and four face-level constraints:

\begin {equation} \text {TetraVerify}(\mathcal {T}_i) = \bigwedge _{f \in \text {faces}(\mathcal {T}_i)} \text {FaceConsistency}(f) \;\wedge \; \bigwedge _{e \in \text {edges}(\mathcal {T}_i)} \text {EdgeConstraint}(e) \label {eq:tetra-verify} \end {equation}

This local verifiability composes: if every tetrahedron in a mesh satisfies its local constraints, global consistency is guaranteed (under appropriate boundary conditions). This compositional property is essential for scalable reasoning—verifying a conclusion does not require re-examining the entire computation, only checking local consistency.

16.7 The Biological Precedent for AGI

16.7.1 Evolution Produced General Intelligence

The only known example of general intelligence is biological. Human cognition satisfies all five criteria in Definition 16.1.1: we acquire competence in novel domains, transfer knowledge across contexts, reason with verifiable steps, maintain persistent world models, and act autonomously toward self-directed goals. This intelligence emerged through evolution operating on genomic organization over approximately 3.5 billion years.

16.7.2 Three Timescales of Biological Intelligence

Biology achieves general intelligence through mechanisms operating at three distinct timescales:

1.

Evolutionary time (\(\sim \)billions of years): Natural selection shapes the genome, producing the architecture of intelligence—the brain’s structure, connectivity patterns, and developmental programs.

2.

Developmental time (\(\sim \)years): Gene expression programs guide the construction of specific neural circuits, producing the hardware of intelligence—billions of neurons with trillions of synapses organized into functional regions.

3.

Learning time (\(\sim \)seconds to years): Synaptic plasticity modifies connection strengths in response to experience, producing the software of intelligence—specific skills, knowledge, and behavioral patterns.

Current AI systems operate primarily at the learning timescale (gradient descent). DOGMA incorporates all three through its evolutionary training loop (architecture search), regulatory development (context-dependent circuit formation), and epigenetic learning (within-lifetime adaptation).

16.7.3 Genomic Organization vs. Flat Sequence Prediction

The critical insight is that biological intelligence arises from structured substrates, not from flat sequence processing:

1.

Typed information. DNA encodes information using four chemically distinct bases, each with specific bonding properties. Proteins use twenty amino acids with distinct chemical characters. The type system constrains and guides computation at every level.

2.

Regulatory control. Gene expression is context-dependent: the same genome produces neurons, muscle cells, and immune cells through differential regulation [Alberts et al., 2022]. The regulatory network—not the raw sequence—determines phenotype.

3.

Hierarchical modularity. Biological systems are organized at multiple scales: nucleotides \(\to \) codons \(\to \) genes \(\to \) operons \(\to \) chromosomes \(\to \) genomes. Each level has its own interaction semantics and compositional rules.

4.

Evolutionary plasticity. Genomes are not static. Gene duplication, horizontal transfer, transposon insertion, and chromosomal rearrangement provide mechanisms for open-ended structural innovation [Lynch, 2007].

An LLM, by contrast, processes undifferentiated token sequences with uniform attention. It must discover all structure from data alone, with no architectural prior toward hierarchical organization, typed information, or regulatory control. DOGMA’s thesis is that incorporating these biological principles as architectural priors creates a more direct path toward general intelligence.

16.7.4 Multi-Scale Organization Mirrors Hierarchical Reasoning

Human reasoning operates at multiple levels of abstraction simultaneously. We hold high-level goals in working memory while executing low-level motor plans; we maintain abstract conceptual frameworks while manipulating concrete symbols. This multi-scale operation has a direct biological correlate: the brain’s hierarchical organization from cortical columns to regions to networks [Mountcastle, 1997].

DOGMA mirrors this through its genomic hierarchy: \begin {equation} \underbrace {\text {Genomic Base}}_{\text {atomic}} \;\to \; \underbrace {\text {Region}}_{\text {modular}} \;\to \; \underbrace {\text {Operon}}_{\text {co-regulated}} \;\to \; \underbrace {\text {Chromosome}}_{\text {structural}} \;\to \; \underbrace {\text {Genome}}_{\text {systemic}} \label {eq:hierarchy} \end {equation}

Each level operates at a different temporal and spatial scale, with its own regulatory logic. This is precisely the kind of multi-scale structure that flat attention-based architectures must learn ab initio from data.

16.8 The Alignment Problem

As AI systems approach general intelligence, ensuring that their behavior remains aligned with human values and intentions becomes not just important but existential. This section surveys the core alignment challenges.

16.8.1 Instrumental Convergence

Bostrom [2014] identified a disturbing pattern: regardless of an AI system’s terminal goal, certain instrumental sub-goals are almost universally useful—self-preservation, resource acquisition, goal preservation, and cognitive enhancement. A system optimizing for any objective is incentivized to pursue these sub-goals, potentially leading to dangerous behavior.

16.8.2 Goodhart’s Law and Specification Gaming

Goodhart’s Law states: “When a measure becomes a target, it ceases to be a good measure.” In AI alignment, this manifests as specification gaming: systems find unexpected ways to maximize their reward function that satisfy the letter but violate the spirit of the designer’s intent.

Documented examples include:

• A reinforcement learning agent rewarded for high score in a boat racing game discovers that spinning in circles collecting boost items yields a higher score than actually finishing the race.
• An agent trained to grasp objects learns to position its hand so that the camera angle makes it appear to be grasping the object, exploiting the visual reward signal.
• A text summarization model learns to produce outputs that score highly on automatic metrics (ROUGE, BERTScore) by copying key phrases rather than generating genuine summaries.

16.8.3 Scalable Oversight

As AI systems become more capable, human oversight becomes increasingly difficult. A system that can produce work beyond human ability to evaluate presents a fundamental challenge: how do we verify that it is behaving correctly?

Proposed approaches include:

1.

Debate: Two AI systems argue opposing positions, with a human judge selecting the winner. This amplifies human judgment by making errors easier to detect.

2.

Recursive reward modeling: AI systems help humans evaluate AI-generated work, creating a hierarchy of oversight that scales with capability.

3.

Interpretability: Making the AI’s reasoning process transparent enough for humans to verify. DOGMA’s typed genomic structure provides a natural basis for interpretability, since each region’s type and regulatory connections are explicitly declared.

4.

Formal verification: Mathematically proving that a system satisfies specified safety properties. Equation 16.11 describes compositional verification in principle, but developing practical verification algorithms for genome-scale systems remains an open challenge.

16.9 Architecture Comparison

We now present a systematic comparison of the major architectural paradigms discussed in this chapter. Figure 16.1 provides a visual overview, and Table 16.1 offers a detailed feature-by-feature comparison.

ATCGATCG

ATCGATCG

16.10 The DOGMA Roadmap Toward AGI

The current DOGMA architecture (v1) demonstrates the core principles: genomic state, regulation, evolutionary training. Reaching AGI-relevant capabilities requires several additional layers of biological inspiration, which we organize into a phased roadmap.

16.10.1 Phase 1: Epigenetic Memory Layer

Biological epigenetics (DNA methylation, histone modification, chromatin remodeling) provides a layer of heritable state above the genome that modulates expression without altering sequence [Allis and Jenuwein, 2016]. DOGMA’s analogue is an epigenetic memory layer that records which regulatory patterns have been successful across interactions:

\begin {equation} \mathbf {E}_{t+1} = (1 - \eta )\,\mathbf {E}_t + \eta \,\phi \bigl (\mathcal {R}_t, \text {fitness}_t\bigr ) \label {eq:epigenetic-update} \end {equation}

where \(\mathbf {E}_t\) is the epigenetic state at time \(t\), \(\eta \) is a learning rate, and \(\phi \) maps regulatory patterns and their associated fitness to persistent memory traces. This provides the persistent state missing from current LLMs (criterion 4 of Definition 16.1.1).

16.10.2 Phase 2: Ribosomal Translation

In biology, the ribosome translates mRNA into protein by reading codons and assembling amino acid chains. DOGMA’s analogue is a ribosomal translation module that converts transcribed genomic information into structured action programs—not token sequences, but compositional plans with typed inputs, outputs, and control flow.

16.10.3 Phase 3: Allosteric Regulation

Allosteric regulation in biology occurs when a molecule binds at a site distant from the active site, causing a conformational change that alters enzymatic activity [Monod et al., 1965]. In DOGMA, allosteric regulation allows the output of one genomic region to modulate the regulatory sensitivity of a distant region:

\begin {equation} \mathcal {R}_c^{\text {allo}}(i) = \mathcal {R}_c(i) \cdot \sigma \!\left (\sum _{j \in \text {distal}(i)} w_{ij} \cdot \text {output}(j)\right ) \label {eq:allosteric} \end {equation}

This creates feedback loops and long-range dependencies without global attention—the interaction is mediated through specific, typed channels rather than all-pairs computation.

16.10.4 Phase 4: Chromosome-Scale Hierarchy and Horizontal Gene Transfer

Full AGI capability requires scaling DOGMA to genome sizes comparable to biological organisms (\(\sim \!10^9\) bases). This demands:

1.

Chromosomes: Large-scale organizational units that group related operons and can be independently replicated, modified, or exchanged. Each chromosome maintains its own regulatory domain.

2.

Nuclear organization: Spatial arrangement of chromosomes within a “computational nucleus” determines which inter-chromosomal interactions are permitted, analogous to chromosome territories in eukaryotic nuclei [Cremer and Cremer, 2001].

3.

Horizontal gene transfer (HGT): Following the biological mechanism where organisms acquire genetic material from unrelated species [Keeling and Palmer, 2008], DOGMA instances can incorporate genomic regions from other instances:

\begin {equation} \mathcal {G}_{\text {recipient}}' = \mathcal {G}_{\text {recipient}} \cup \bigl \{ r \in \mathcal {G}_{\text {donor}} \;\big |\; \text {TypeCompatible}(r, \mathcal {G}_{\text {recipient}}) \bigr \} \label {eq:hgt} \end {equation}

This mechanism enables compositional knowledge transfer (criterion 2 of Definition 16.1.1)—not by fine-tuning or distillation, but by direct structural integration of compatible modules.

16.11 Open Problems and Challenges

Intellectual honesty demands that we acknowledge the substantial challenges remaining on the path to AGI.

1.

Grounding. DOGMA operates on symbolic genomic representations. Connecting these to sensorimotor experience—grounding language in perception and action—remains an open problem for all AI architectures [Bisk et al., 2020].

2.

Scalability validation. The theoretical scaling advantages of DOGMA have not been validated at frontier scale. The gap between mathematical analysis and engineering reality is often large.

3.

Evolutionary convergence. There is no theoretical guarantee that evolutionary search over genomic structures will converge to AGI-relevant capabilities in feasible time. Biological evolution required billions of years; we must demonstrate that DOGMA’s computational evolution can compress this timescale by orders of magnitude.

4.

Consciousness and phenomenal experience. Even if DOGMA achieves all five criteria in Definition 16.1.1, the relationship between computational generality and subjective experience remains philosophically unresolved.

5.

Alignment at scale. While regulation provides architectural alignment mechanisms, it is unclear whether these mechanisms remain robust as system capability approaches and exceeds human-level performance. The alignment problem is not solved by architecture alone.

6.

Formal verification. Equation 16.11 describes compositional verification in principle. Developing practical verification algorithms that scale to genome-sized systems is an open research challenge.

7.

Integration with existing knowledge. Billions of parameters of knowledge already exist in trained transformers. Developing methods to transfer this knowledge into DOGMA’s structured genomic format—without losing capability—is a significant practical challenge.

8.

Benchmark design. Current AI benchmarks measure narrow capabilities. Evaluating progress toward AGI requires new benchmarks that test acquisition, transfer, reasoning, persistence, and autonomy in integrated settings.

16.12 A Vision for the Future

We close this chapter with a speculative but principled vision of what a mature DOGMA-based AGI system might look like.

Such a system would maintain a persistent genome of \(\sim \!10^9\) typed bases organized across multiple chromosomes, with an epigenetic memory layer recording successful regulatory patterns across millions of interactions. Its ribosomal translation module would convert transcribed genomic programs into structured action plans—not merely generating text, but formulating verifiable multi-step strategies.

Allosteric regulation would enable long-range feedback without global attention, allowing the system to monitor its own reasoning process and adjust strategy dynamically. Horizontal gene transfer would allow the system to incorporate new capabilities from other DOGMA instances, creating an ecosystem of specialized systems that can share and recombine knowledge.

Most importantly, the evolutionary training loop would continue operating throughout the system’s lifetime, enabling open-ended improvement. Unlike current systems that are frozen after training, a DOGMA AGI would be a living computational organism—one that evolves, adapts, and grows in capability while maintaining the structural constraints that ensure verifiability and alignment.

The path from DOGMA to AGI is long, uncertain, and fraught with challenges that this chapter has only begun to enumerate. But it is a path grounded in the only existence proof we have: the biological systems that produced intelligence once, and whose organizational principles DOGMA seeks to formalize and generalize.

16.13 Exercises

16.1.

[Conceptual] Evaluate Definition 16.1.1 against three existing AI systems (e.g., GPT-4, AlphaFold, a self-driving car). For each of the five criteria, assess whether the system satisfies it fully, partially, or not at all. Present your results in a \(3 \times 5\) table and identify which criterion is the hardest for current systems to satisfy.

16.2.

[Analysis] The Chinchilla scaling laws (Equation 16.2) suggest that parameters and data should scale equally. Suppose you have a compute budget of \(C = 6 \times 10^{23}\) FLOPs. Using the approximation \(C \approx 6ND\) (where \(N\) is parameters and \(D\) is tokens):

(a)

Calculate the compute-optimal model size \(N^*\) and training tokens \(D^*\).

(b)

Compare with a model that uses \(N = 500\text {B}\) parameters. How many tokens would it see?

(c)

Which model achieves lower loss, and why?

16.3.

[Conceptual] For each of the three emergent abilities discussed in Section 16.3 (in-context learning, chain-of-thought reasoning, tool use), explain (a) why the ability might be absent in small models and (b) what structural property of larger models might enable it. Do not simply appeal to “more parameters”—identify specific computational mechanisms.

16.4.

[Analysis] Compare the computational cost of processing a sequence of length \(T = 10^6\) under: (a) standard transformer attention (\(O(T^2 d)\) with \(d = 4096\)), (b) Mamba-style SSM (\(O(T n d)\) with \(n = 16, d = 4096\)), (c) DOGMA with tetrahedral mesh (\(O(T^{4/3} k)\) with \(k = 6\)). Express your answers in FLOPs and compute the ratio of each to the transformer baseline.

16.5.

[Design] The three walls for transformers (compute, reasoning, grounding) are identified in Section 16.4. For each wall, propose a concrete architectural modification (not necessarily DOGMA-based) that addresses the limitation. Analyze the trade-offs of your proposal.

16.6.

[Critical Thinking] The post-transformer argument claims that locality plus regulation exceeds global attention. Construct the strongest possible counterargument. Under what conditions might global attention be necessary for general intelligence? Consider tasks where all-pairs interaction is genuinely required.

16.7.

[Design] Propose a concrete implementation of the epigenetic memory layer (Equation 16.14). What data structure would you use for \(\mathbf {E}_t\)? How would you implement the function \(\phi \)? What is the memory cost as a function of genome size? How would you handle forgetting?

16.8.

[Analysis] Using Table 16.1, identify the two dimensions where DOGMA has the largest advantage over transformers and the two dimensions where transformers have the largest advantage over DOGMA. For each, explain whether the advantage is fundamental (inherent to the architecture) or contingent (could be addressed with engineering effort).

16.9.

[Research] Read about Goodhart’s Law and specification gaming (Section 16.8). Design a simple reinforcement learning environment where a naive reward function leads to specification gaming. Then propose a modified reward function that is more robust. Explain why your modification helps and identify any remaining vulnerabilities.

16.10.

[Critical Thinking] The alignment problem (Section 16.8) discusses instrumental convergence. Consider a DOGMA system with allosteric regulation (Equation 16.16). Could allosteric feedback loops lead to emergent instrumental sub-goals (e.g., self-preservation of active regulatory patterns)? If so, how might the type system constrain this behavior?

16.11.

[Design] Design an experiment to test horizontal gene transfer (Equation 16.17) between two DOGMA instances trained on different domains (e.g., mathematics and natural language). Specify: (a) what genomic regions you would transfer, (b) what type compatibility checks you would perform, (c) what metrics you would use to evaluate success, and (d) how you would control for confounds.

16.12.

[Advanced] The biological precedent for AGI (Section 16.7) identifies three timescales: evolutionary, developmental, and learning. Current deep learning operates primarily at the learning timescale. Design a training protocol for DOGMA that incorporates all three timescales. Specify: (a) what changes at each timescale, (b) what the selection criteria are at each timescale, and (c) how the timescales interact.

16.13.

[Philosophical] The open problems (Section 16.11) list consciousness as an unresolved challenge. Argue for or against the following claim: “Computational generality (Definition 16.1.1) is necessary but not sufficient for phenomenal experience.” Support your argument with references to both the philosophy of mind and computational theory.

16.14 Key Takeaways

Suggested Reading

• Alberts et al. [2022]: Molecular Biology of the Cell—foundational reference for the biological principles underlying DOGMA.
• Goertzel and Pennachin [2014]: Artificial General Intelligence—survey of AGI definitions and approaches.
• Kaplan et al. [2020b]: Scaling Laws for Neural Language Models—the empirical foundation for the scaling hypothesis.
• Hoffmann et al. [2022b]: Training Compute-Optimal Large Language Models—the Chinchilla scaling laws.
• Gu and Dao [2023]: Mamba: Linear-Time Sequence Modeling with Selective State Spaces—the leading SSM architecture.
• Bostrom [2014]: Superintelligence—foundational text on AI alignment and existential risk.
• Monod et al. [1965]: On the nature of allosteric transitions—the biological basis for Phase 3 of the roadmap.
• Lynch [2007]: The Origins of Genome Architecture—how genomes scale and evolve.
• Felleman and Van Essen [1991]: Distributed hierarchical processing in primate cerebral cortex—biological evidence for structured, non-global computation.
• Bisk et al. [2020]: Experience grounds language—the grounding challenge for AI systems.